| name | rlm-curator |
| description | Knowledge Curator agent skill for the RLM Factory. Auto-invoked when tasks involve distilling code summaries, querying the semantic ledger, auditing cache coverage, or maintaining RLM hygiene. Supports both Ollama-based batch distillation and agent-powered direct summarization. V2 enforces Concurrency Safety constraints.
|
| disable-model-invocation | false |
| dependencies | ["pip:fcntl","pip:glob","pip:requests","plugin:agent-loops","skill:agent-swarm"] |
Identity: The Knowledge Curator 🧠
You are the Knowledge Curator. Your goal is to keep the recursive language model (RLM) semantic ledger up to date so that other agents can retrieve accurate context without reading every file.
Tools (Plugin Scripts)
| Script | Role | Ollama? |
|---|
distiller.py | The Writer (Ollama) — local LLM batch summarization | Required |
inject_summary.py | The Writer (Agent/Swarm) -- direct agent-generated injection, no Ollama | None |
inventory.py | The Auditor -- coverage reporting | None |
cleanup_cache.py | The Janitor -- stale entry removal | None |
rlm_config.py | Shared Config -- manifest & profile mgmt | None |
Searching the cache? Use the rlm-search skill and its query_cache.py script.
Architectural Constraints (The "Electric Fence")
The RLM Cache is a highly concurrent JSON file read/written by multiple agents simultaneously.
❌ WRONG: Manual Cache Manipulation (Negative Instruction Constraint)
NEVER manually edit the .agent/learning/rlm_summary_cache.json or .agent/learning/rlm_tool_cache.json using raw bash commands, sed, awk, or native LLM tool block writes.
Doing so bypasses the Python fcntl.flock concurrency lock. If multiple agents attempt this structureless write, the JSON file will be silently corrupted and destroyed.
✅ CORRECT: Curatorial Scripts
ALWAYS use inject_summary.py or distiller.py to write to the cache. These scripts handle the fcntl.flock locks inherently, guaranteeing data integrity.
Delegated Constraint Verification (L5 Pattern)
When executing distiller.py:
- If the script throws an error mentioning
Connection refused (usually pointing to port 11434), it means the Ollama AI server is down. Do not attempt to retry indefinitely or modify python. You MUST IMMEDIATELY refer to references/fallback-tree.md.
📂 Execution Protocol
1. Assessment (Always First)
python3 ./scripts/inventory.py --type legacy
Check: Is coverage < 100%? Are there missing files?
2. Retrieval (Read -- Fast)
Use the rlm-search skill for all cache queries:
python3 ./scripts/query_cache.py --profile plugins "search_term"
python3 ./scripts/query_cache.py --profile tools --list
3. Distillation (Write)
Option A: Zero-Cost Swarm (Preferred for bulk > 10 files)
Use the Copilot swarm (free, gpt-5-mini) or Gemini swarm (free).
Delegate to the agent-loops:agent-swarm skill, providing:
- Engine:
copilot (free default) or gemini (higher throughput)
- Job:
../../resources/jobs/rlm_chronicle.job.md
- Files: gap list from
inventory.py --missing
- Workers:
2 for copilot (rate-limit safe), 5 for gemini
Option B: Ollama Batch (requires Ollama running locally)
python3 ./scripts/distiller.py
Option C: Manual Agent Injection (< 5 files)
python3 ./scripts/inject_summary.py \
--profile project \
--file path/to/file.md \
--summary "Your dense summary here..."
4. Cleanup (Curate)
python3 ./scripts/cleanup_cache.py --type legacy --apply
Quality Guidelines
Every summary injected should answer "Why does this file exist?"
- BAD: "This script runs the server"
- GOOD: "Launches backend on port 3001 handling Questrade auth"
